Explicit Policy

Categorical Missing Uses the Mode

The categorical replacement is the exact observed mode.

highlighted = computed this step

Use only observed colors

The observed colors are red, blue, and red. The missing color is not counted when choosing the mode.

observed colors=(red,blue,red)\text{observed colors}=(\text{red},\text{blue},\text{red})
Missing Values ExactlyExact missing-cell replacements from a pinned policy.explicit missing-value policyrowage rawcolor rawage filledcolor filledage_missingcolor_missingA2red2red00B?blue6blue10C6?6red01D10red10red00age mean: (2+6+10)/3 = 6color mode: red count 2, blue count 1add missingness indicatorsexplicit preprocessing policy for displayed missing cellsnumeric uses observed mean; categorical uses observed modemissingness indicators stay visibleNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth

Count the exact mode

Red appears 2 times, and blue appears 1 time. The pinned mode is red.

red count=2,blue count=1\text{red count}=2,\quad \text{blue count}=1
Missing Values ExactlyExact missing-cell replacements from a pinned policy.explicit missing-value policyrowage rawcolor rawage filledcolor filledage_missingcolor_missingA2red2red00B?blue6blue10C6?6red01D10red10red00age mean: (2+6+10)/3 = 6color mode: red count 2, blue count 1add missingness indicatorsexplicit preprocessing policy for displayed missing cellsnumeric uses observed mean; categorical uses observed modemissingness indicators stay visibleNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth

Pin the categorical policy

The categorical policy is explicit: color missing uses the observed mode red.

color ? uses red\text{color } ? \text{ uses red}
Missing Values ExactlyExact missing-cell replacements from a pinned policy.explicit missing-value policyrowage rawcolor rawage filledcolor filledage_missingcolor_missingA2red2red00B?blue6blue10C6?6red01D10red10red00age mean: (2+6+10)/3 = 6color mode: red count 2, blue count 1add missingness indicatorsexplicit preprocessing policy for displayed missing cellsnumeric uses observed mean; categorical uses observed modemissingness indicators stay visibleNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth